benchgap
Calibration

Hermes Bench → CWE-Bench-AA

CWE-Bench-AA is estimated from Hermes Bench with a Michaelis–Menten curve fitted on 10 models measured on both: y = 2.0000·x / (1.73862 + x), R² = 0.36, cross-validated error 4.6 pp. It is used for 2 estimates.

Estimated modelHermes BenchCWE-Bench-AASource
Hy4 preview70.8%57.9%estimated ± 4.6 pp, medium confidence
Ling 3.0 Flash53.8%47.3%estimated ± 4.6 pp, low confidence

Fitted on

ModelHermes BenchCWE-Bench-AA
Claude Opus 5.576.7%58.3%
Gemini 3.8 Flash62.2%52.5%
GLM-5.371.8%55.8%
GLM-5.3-Flash68.9%55.8%
GPT-6 Astra73.7%63.3%
GPT-6 Luna70.5%56.7%
GPT-6 Sol73.5%64.2%
Grok 4.774.4%68.3%
Kimi K371.5%53.3%
Step 5 Preview69.2%52.5%